The Architecture of the FAIM-1 Symbolic Multiprocessing System

Alan L. Davis, Shane V. RobisonIJCAI 1985

Introduction

In the mid-1980s, symbolic AI applications—characterized by logic programming, object-oriented structures, and intensive pattern matching—faced a significant performance barrier. Traditional Von Neumann architectures were optimized for sequential numeric tasks, leading to a "bottleneck" where processors spent more time managing memory than executing AI logic. The FAIM-1 (Fanatic AI Machine-1), developed by Schlumberger Palo Alto Research, was designed as an ultra-concurrent symbolic multiprocessor to overcome these limitations, targeting a 100-fold performance increase over contemporary workstations.

System Architecture

The FAIM-1 architecture is built upon three integrated pillars: a hybrid programming language, a specialized processing element, and a scalable interconnection topology.

The OIL Language

"Our Intermediate Language" (OIL) serves as the linguistic foundation, blending logic programming (Prolog-style), object-oriented programming, and procedural code (Lisp-style). This multi-paradigm approach allows developers to select the most efficient style for specific tasks, avoiding the overhead of emulating one paradigm within another.

The Hectogon Processing Element

The fundamental building block of the system is the Hectogon, a processing element composed of six specialized subsystems designed to offload complex tasks from the main evaluator:

  • FRISC (Fanatically Reduced Instruction Set Computer): The primary evaluation engine.
  • ISM (Instruction Stream Memory): An intelligent memory subsystem that performs branch prediction and instruction prefetching, effectively hiding control flow overhead.
  • CxAM (Context Addressable Memory): Hardware-accelerated associative memory for rapid pattern matching.
  • SPUN (Streamed Pipelined UNifier): Dedicated hardware for logic unification.
  • SRAM & Post Office: Specialized structure RAM for object management and an autonomous controller for inter-processor communication.
The Hectogon processing element and its interconnection within the hexagonal mesh.

Interconnection Topology

Hectogons are arranged in a hexagonal mesh that employs a "twisted toroidal" wrap-around. This 3-axis topology ensures an optimal switching diameter, minimizing the communication distance between any two processors in a large-scale system.

A large-scale hexagonal tiling of Hectogons forming a processing surface.

Resource Allocation and Scalability

To manage the inherent concurrency of AI programs, the FAIM-1 utilizes a static "Allocator." This tool employs simulated annealing to map program tasks onto the physical hexagonal grid, aiming to minimize inter-processor communication latency.

The system's modular "tiling" design allows it to scale to an arbitrary number of processors. The researchers initially developed a 19-processor prototype, known as an "E-3 surface," demonstrating the feasibility of the VLSI-based design. By moving intelligence into the memory subsystems (e.g., CxAM and ISM), the FRISC processor is freed to focus entirely on computation, significantly boosting throughput.

Conclusion and Future Outlook

The FAIM-1 represents a radical departure from conventional computer design by embedding AI-specific structures directly into hardware. While the system demonstrated significant potential for symbolic processing, the authors noted that highly dynamic AI workloads might eventually require runtime dynamic load balancing. Furthermore, the transition from sequential to concurrent programming remains a significant challenge for developers. Despite these hurdles, the FAIM-1's architecture—particularly its use of intelligent memory and optimized topologies—provided a blueprint for high-performance symbolic multiprocessing.

Summary updated

BibTeX
@inproceedings{davis1985architecture,
  author = {Alan L. Davis and Shane V. Robison},
  title = {{The Architecture of the FAIM-1 Symbolic Multiprocessing System}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1985},
}
Figure